---
title: "一個免費的 MCP 讀取 212 個遺留檔案並在任何代理接觸程式碼之前撰寫規格"
url: "https://reamcp.dev/zh-tw/cases/a-free-mcp-reads-212-legacy-files-and-writes-the-spec-before-831c92"
source: "https://x.com/i/status/2111530936441446417"
source_type: "x"
author: "Primee32"
author_handle: "Primee32"
stars: 0
category: "應用功能重現"
categories: ["app-recreation"]
tags: ["code-rewrite", "legacy-modernization", "mcp", "reverse-engineering", "spec-generation", "verification"]
added: "2026-10-19T02:06:48.267Z"
updated: "2026-10-17T18:53:00.000Z"
locale: "zh-tw"
content_locale: "zh-tw"
---

# 一個免費的 MCP 讀取 212 個遺留檔案並在任何代理接觸程式碼之前撰寫規格

一篇 X 貼文描述了一個工作流程：REA (morluto/rea) 將遺留計費系統逆向工程為 spec.md（14 個端點、37 條規則、9 個邊緣情況），十二個 Sonnet 代理從規格切片建置模組，REA VERIFY 針對遺留系統重播新系統（148/148 行為匹配）。貼文連結到 MIT 授權的 GitHub 專案，並給出了 npx rea-agents setup 安裝指令，以及關於將工具結果傳送給模型提供者的警告。

## Description

一個免費的 MCP 讀取 212 個遺留檔案並在任何代理接觸程式碼之前撰寫規格

大多數重寫都失敗在同一個地方。你把整個舊系統丟進一個模型，它就被淹沒了

這個設定將工作拆分：

OPUS 5.5 > 撰寫計畫，然後在剩餘執行中閒置
REA > 一次性逆向工程 legacy/billing，並將其轉換為 spec.md，包含 14 個端點、37 條規則、9 個邊緣情況
SONNET 5.5 x12 > 每個一個模組，從規格的約 1.5k token 切片建置。它們都不開啟舊程式碼
REA VERIFY > 針對遺留系統重播新系統，148 個行為中有 148 個匹配

在螢幕上，群組在 7 分 05 秒內完成，花費 $0.88。opus $0.22，REA $0.32，十二個 sonnet $0.34。僅 Opus 5.5 就處於 23% 和 $1.20，仍在執行

影片將該比較標記為說明性，因此將數字視為演示

REA 不來自 Anthropic。它是 GitHub 上的 morluto/rea，MIT 授權且免費，你只需為模型付費。一行安裝它：npx rea-agents setup

在你將其指向客戶程式碼之前：
> REA 在本機分析，但你的代理仍將工具結果傳送給模型提供商
> 來自私有應用的解編譯程式碼最終會進入 Anthropic 或 OpenAI，受其資料政策約束

它今天是免費的。在有人將其作為 $500 課程出售之前儲存它

## Original README

A FREE MCP READS 212 LEGACY FILES AND WRITES THE SPEC BEFORE ANY AGENT TOUCHES THE CODE

most rewrites go wrong in the same place. you dump the whole old system into one model and it drowns

this setup splits the work:

OPUS 5.5 > writes the plan, then sits idle for the rest of the run
REA > reverse-engineers legacy/billing once and turns it into spec.md with 14 endpoints, 37 rules, 9 edge cases
SONNET 5.5 x12 > one module each, built from a ~1.5k token slice of the spec. none of them opens the old code
REA VERIFY > replays the new system against the legacy one, 148 of 148 behaviours match

on screen the swarm finishes in 7m 05s for $0.88. opus $0.22, REA $0.32, twelve sonnets $0.34. Opus 5.5 alone sits at 23% and $1.20, still running

the video marks that comparison as illustrative, so take the numbers as a demo

REA doesn't come from Anthropic. it's morluto/rea on github, MIT licensed and free, you only pay for the model. one line installs it: npx rea-agents setup

before you point it at client code:
> REA analyzes locally, but your agent still sends tool results to the model provider
> decompiled code from a private app ends up with Anthropic or OpenAI, under their data policy

it's free today. save it before someone sells it as a $500 course